Vacuum coating process parameter self-adaptive adjusting method and system

Abstract

The invention discloses a vacuum coating process parameter adaptive adjustment method and system, the system comprises a data acquisition layer, a DRL intelligent agent and an execution layer, the DRL intelligent agent comprises a strategy network and an evaluation network, the DRL intelligent agent adopts a double-loop learning mechanism, and the DRL intelligent agent comprises an offline imitation learning module and an online reinforcement learning module. According to the technical scheme, through DRL decoupling multi-roller coupling, the tension fluctuation control precision is obviously improved compared with traditional PID control; factors such as roller diameter thermal deformation and environmental disturbance are automatically compensated through an online learning mechanism, the dynamic adaptability is enhanced, and manual adjustment is not needed; in addition, closed-loop control of visual defect detection is fused in the technical scheme, so that the film thickness uniformity is greatly improved, and the wrinkle defect rate is obviously reduced.

Description

Technical Field

[0001] This application relates to, in particular, a method and system for adaptive adjustment of vacuum coating process parameters. Background Technology

[0002] With the explosive growth of the global new energy industry, the market demand for power batteries, as key components in core areas such as new energy vehicles and energy storage systems, continues to rise. Among them, high-quality lithium batteries, with their advantages of high energy density and long cycle life, have become the core focus of industry competition. The improvement of lithium battery performance is highly dependent on the preparation process of positive and negative electrode base film materials. In order to achieve effective bonding between electrodes and current collectors and reduce interfacial impedance, aluminum foil or copper foil current collectors need to be precisely coated on both sides of the PET base film. Moreover, the uniformity of the coating thickness and the stability of the base film tension directly determine the charging and discharging efficiency and safety performance of the battery, which places stringent requirements on the vacuum coating process.

[0003] In high-end lithium battery production, continuous roll-to-roll coating of ultra-wide (width greater than 1.5 meters) positive and negative electrode thin film materials has become the mainstream process. This process requires the base film to be coated with aluminum or copper on both sides in a single operation through evaporation coating technology in a vacuum environment. The equipment structure and control logic are extremely complex: to ensure the flattening effect of the ultra-wide film, at least 10 power rollers are required (the more rollers and the smaller the roller spacing, the better the flattening effect). In actual production, an 11-roll layout is usually adopted (including unwinding roller, flattening roller, partition roller, traction roller, and rewinding roller), and each roller needs to achieve dynamic tension balance. At the same time, the real-time matching of evaporation source power and winding speed directly affects the uniformity of coating thickness. The fluctuations in temperature, humidity, and vacuum degree in the vacuum chamber, as well as the dynamic disturbances such as roller diameter changes and friction coefficient drift caused by thermal expansion and contraction of the traction roller, further increase the difficulty of process control.

[0004] Currently, the industry's control solutions for vacuum coating processes have significant limitations, making it difficult to meet the production demands for ultra-wide widths and high precision. These limitations are mainly reflected in the following three aspects:

[0005] Firstly, the challenge of multi-roller tension coupling control is significant. Traditional control schemes often employ PID control algorithms, but these can only achieve simple tension adjustment of a single roll and cannot address the dynamic coupling effect between multiple traction rolls. In an 11-roller continuous transport system, there is significant interference in tension transmission between the rolls. PID control is prone to tension oscillations (measured tension fluctuations can reach ±10%), leading to localized stress concentration in the film and resulting in defects such as wrinkles and tears, severely impacting product yield. For example, the "High-Speed ​​Constant Tension Equipment and Control Method for Vacuum Coating" disclosed in patent CN201510729302 is only suitable for single-sided, small-width coating applications with low control precision requirements. Its single-roller constant tension control logic does not consider the oscillation problem caused by multi-roller coupling, cannot guarantee coordinated operation between rolls, and is unsuitable for the flattening and tension requirements of ultra-wide films.

[0006] Secondly, the matching efficiency of process parameters is low and the response is lagging. The matching relationship between the evaporation source power and the winding speed directly determines the coating thickness. In current production, this relies heavily on manual adjustments based on the operator's experience: when a film thickness deviation is detected, the operator needs to judge the amount of power or speed adjustment based on experience. This process has a significant response delay, and the adjustment accuracy is greatly affected by the operator's skill level. In addition, traditional solutions use fixed process parameter formulas. For different base film materials, widths, or target film thicknesses, the machine needs to be stopped for parameter readjustment, which not only extends the production cycle but also increases trial and error costs.

[0007] Third, it has poor adaptability to dynamic operating conditions. The vacuum coating production process involves various dynamic disturbances: first, changes in the physical properties of the roller system, such as the difference in thermal expansion coefficients between metal traction rollers and rubber flattening rollers leading to changes in roller diameter, and fluctuations in the surface friction coefficient caused by the operation of the evaporation roller cooling system; second, fluctuations in environmental parameters, such as instantaneous changes in temperature, humidity, and vacuum level within the vacuum chamber affecting the evaporation rate and the physical properties of the film; and third, batch-to-batch differences in materials, with slight variations in the mechanical strength and surface tension of base films from different batches. Existing control schemes use fixed parameter modes, which cannot detect and compensate for these dynamic disturbances in real time, resulting in poor coating thickness uniformity and a high surface defect rate.

[0008] Therefore, there is an urgent need to develop a DRL control scheme that is adapted to the characteristics of vacuum coating process in order to break through the bottleneck of traditional technology. Summary of the Invention

[0009] In view of the above problems, this application provides an adaptive adjustment method and system for vacuum coating process parameters to solve the technical problem of coordinated control of ultra-wide film flattening and tension.

[0010] To achieve the above objectives, this application provides an adaptive adjustment system for vacuum coating process parameters, comprising:

[0011] The main body of the vacuum coating production line has multiple power rollers arranged sequentially along the base film transport path, including unwinding rollers, flattening rollers, partition rollers, traction rollers and take-up rollers;

[0012] The data acquisition layer includes tension sensors installed between each roller, a sheet resistance detector for detecting film thickness, a CCD vision system for identifying surface defects, and temperature and humidity sensors and vacuum sensors for acquiring environmental parameters of the vacuum chamber.

[0013] The decision layer includes a DRL agent, which comprises a policy network and an evaluation network, and its input is a state vector S fused from multimodal data. t The output is the process parameter adjustment action vector A. t The DRL agent employs a dual-loop learning mechanism, including an offline imitation learning module and an online reinforcement learning module. The offline imitation learning module is configured to: process the expert dataset D... expert Cleaning, normalization, and alignment preprocessing are performed; the policy network π is trained using the behavior cloning algorithm. θ In expert state S t For input, expert action A t For the label, minimize the mean squared error loss function L BC (θ)=E[(π θ (S t ), A t )²], output the initial network parameters θ init ;

[0014] The execution layer includes a servo motor controller connected to each power roller and a power controller connected to the evaporation source. The servo motor controller supports synchronous operation mode and speed loop saturation torque limiting mode.

[0015] A knowledge base is used to store expert datasets D, which consist of historical high-quality process data. expert The dataset contains state-action-reward tuples (S... t A t ,R t ).

[0016] Furthermore, the inter-roller tension collaborative control algorithm is configured in the decision layer, and the algorithm includes:

[0017] Tension transfer model in the unwinding zone: T5 = F1 - ΣF i , i=2 to 5, where T5 is the tension at the input end of the partition roller, F1 is the tension of the unwinding roller, and F2-F5 are the tensions of the flattening roller;

[0018] Tension transfer model in the winding zone: T6=F 11 +ΣF i(i=7 to 10), where T6 is the tension at the output end of the partition roller, F 11 For the tension of the take-up roll, F7-F 10 For the tension of the traction roller;

[0019] Speed ​​gain adjustment rules: Speed ​​gain ΔG of each power roller i The unwinding zone is linearly distributed according to position, and the winding zone uses negative gain. The speed gain of the break roller is 0. The gain coefficient β is calibrated by the feedback value of the tension sensor so that the absolute value of the error between the measured tension and the model calculation value is less than the preset value.

[0020] Furthermore, the online reinforcement learning module is configured as follows:

[0021] Real-time status data is collected periodically, and action A is output through the initialized policy network. t And execute;

[0022] Transfer sample (S) t A t ,R t ,S t+1 Stored in a FIFO-structured experience replay pool;

[0023] Batch data is randomly sampled from the replay pool periodically. The evaluation network parameter φ is updated to minimize the Bellman error, the policy network parameter θ is updated to maximize the expected return, and the target network is updated through soft update rules θ'←τθ+(1-τ)θ' and φ'←τφ+(1-τ)φ', where τ is the soft update coefficient.

[0024] Furthermore, the state vector S t =[F 1..11 D expert ᵣ,T m ,R f N s ,T v H v ,P v ], where F 1..11 For the tension values ​​of 11 power rollers, D r T is the roller diameter parameter. m R is the film surface temperature. f N represents the percentage of the folded area. s T represents the degree of yellowing of the coating. v H v P v These are the vacuum chamber temperature, humidity, and vacuum level, respectively.

[0025] Furthermore, the decision-making layer also includes a reward function module, wherein the reward function is a multi-objective weighted reward function:

[0026] Where ΔF is the change in the standard deviation of tension, ΔH is the deviation between the measured film thickness and the target thickness, and I stable As a stability bonus, the weighting coefficients are w1=0.3, w2=0.4, w3=0.2, and w4=0.1.

[0027] Furthermore, it also includes a safety constraint module, which is configured as follows:

[0028] Action correction layer: Action A output by DRL t Physically limit the tension adjustment to ensure it does not exceed 10% of the current setting.

[0029] Emergency intervention unit: When the percentage of wrinkle area R is detected f When the thickness deviation is >20% or the film thickness deviation |ΔH|>15%, the expert system will take over control.

[0030] Another technical solution of this application is:

[0031] An adaptive adjustment method for vacuum coating process parameters based on the above system includes the following steps:

[0032] S1. Initialize process parameters: Based on the input base film material, width, and target film thickness, match the initial formula from the knowledge base or generate it through expert rules. The formula includes the winding and unwinding tension T. set Main line velocity V roll and evaporation source power P evap ;

[0033] S2. Coordinated calibration of inter-roller tension: After film insertion, the winding and unwinding tension is established. The tension on both sides of the partition roller is calculated through the tension transmission model. The model is calibrated by comparing the actual sensor values ​​and adjusting the speed gain of each roller to balance the tension of each section of film.

[0034] S3. Vacuum environment adaptation: After the base film enters the vacuum chamber, the main line runs at low speed and the speed gain of each roller is finely adjusted to compensate for the tension imbalance caused by the change in the physical properties of the roller;

[0035] S4.DRL Dual-Cycle Optimization: After the vacuum level and evaporation power meet the standards, the main line runs at high speed and coats the film. The DRL intelligent agent collects multi-modal status in real time and outputs action to adjust process parameters after dual-cycle learning.

[0036] S5. Safety monitoring and data archiving: During the production process, the operating status is monitored through the safety constraint module, process parameters are stored in the database in real time, and the formula is added to the knowledge base after production is completed.

[0037] Furthermore, the inter-roller tension collaborative control algorithm is configured in the decision layer, and the algorithm includes:

[0038] Tension transfer model in the unwinding zone: T5 = F1 - ΣFi , i=2 to 5, where T5 is the tension at the input end of the partition roller, F1 is the tension of the unwinding roller, and F2-F5 are the tensions of the flattening roller;

[0039] Tension transfer model in the winding zone: T6=F 11 +ΣF i (i=7 to 10), where T6 is the tension at the output end of the partition roller, F 11 For the tension of the take-up roll, F7-F 10 For the tension of the traction roller;

[0040] Speed ​​gain adjustment rules: Speed ​​gain ΔG of each power roller i The unwinding zone is linearly distributed according to position, and the winding zone uses negative gain. The speed gain of the break roller is 0. The gain coefficient β is calibrated by the feedback value of the tension sensor so that the absolute value of the error between the measured tension and the model calculation value is less than the preset value.

[0041] Furthermore, it also includes the following steps:

[0042] When the percentage of wrinkle area R is detected f When the thickness deviation is >20% or the film thickness deviation |ΔH|>15%, the expert system will take over control.

[0043] Unlike existing technologies, the above-mentioned adaptive adjustment system for vacuum coating process parameters includes a data acquisition layer, a DRL agent, and an execution layer. The DRL agent comprises a policy network and an evaluation network, employing a dual-loop learning mechanism, including an offline imitation learning module and an online reinforcement learning module. This solution decouples multi-roller coupling through DRL, significantly improving tension fluctuation control accuracy compared to traditional PID control. Furthermore, the online learning mechanism automatically compensates for factors such as roller diameter thermal deformation and environmental disturbances, enhancing dynamic adaptability and eliminating the need for manual adjustments. Additionally, this solution integrates closed-loop control with visual defect detection, greatly improving film thickness uniformity and significantly reducing the wrinkle defect rate.

[0044] The above description of the invention is merely an overview of the technical solution of this application. To enable those skilled in the art to better understand the technical solution of this application and to implement it based on the text and drawings, and to make the above-mentioned objectives and other objectives, features, and advantages of this application easier to understand, the following description is provided in conjunction with specific embodiments and accompanying drawings. (See accompanying drawings for details.)

[0045] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on this application.

[0046] In the accompanying drawings of the instruction manual:

[0047] Figure 1 This is a schematic diagram of the overall layout of the vacuum coating system described in the specific implementation method;

[0048] Figure 2 This is a block diagram of the adaptive adjustment system for vacuum coating process parameters described in a specific implementation.

[0049] Figure 3 A flowchart illustrating the adaptive adjustment method for vacuum coating process parameters described in the specific implementation embodiment;

[0050] Figure 4 The following is a flowchart of the DRL dual-loop learning process described in a specific implementation method; Detailed Implementation

[0051] To illustrate the possible application scenarios, technical principles, implementable specific solutions, and achievable objectives and effects of this application in detail, the following description, in conjunction with the listed specific embodiments and accompanying drawings, provides a detailed explanation. The embodiments described herein are merely illustrative of the technical solutions of this application and are therefore intended to limit the scope of protection of this application.

[0052] In this document, the term "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The term "embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment, nor does it specifically limit its independence or connection with other embodiments. In principle, in this application, as long as there are no technical contradictions or conflicts, the technical features mentioned in each embodiment can be combined in any way to form corresponding implementable technical solutions.

[0053] Unless otherwise defined, the technical terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the use of related terms herein is merely for the purpose of describing particular embodiments and is not intended to limit this application.

[0054] In the description of this application, the term "and / or" is used to describe the logical relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A exists, B exists, and A and B exist simultaneously. Additionally, the character " / " in this document generally indicates that the preceding and following objects have an "or" logical relationship.

[0055] In this application, terms such as “first” and “second” are used only to distinguish one entity or operation from another, and do not necessarily require or imply any actual quantity, hierarchy or order relationship between these entities or operations.

[0056] Without further limitations, the use of terms such as “comprising,” “including,” “having,” or other similar open-ended expressions in this application is intended to cover non-exclusive inclusion, which does not exclude the presence of additional elements in a process, method, or product that includes the stated elements, such that a process, method, or product that includes a list of elements may include not only those defined elements but also other elements not expressly listed, or elements inherent to such a process, method, or product.

[0057] As understood in the Examination Guidelines, in this application, expressions such as "greater than," "less than," and "exceeding" are understood to exclude the stated number; expressions such as "above," "below," and "within" are understood to include the stated number. Furthermore, in the description of the embodiments in this application, "multiple" means two or more (including two), and similar expressions related to "multiple" are also understood in this way, such as "multiple groups" and "multiple times," unless otherwise explicitly specified.

[0058] In the description of the embodiments of this application, the space-related expressions used, such as "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "vertical," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," indicate the orientation or positional relationship based on the orientation or positional relationship shown in the specific embodiments or drawings. They are only for the purpose of describing the specific embodiments of this application or for the reader's understanding, and do not indicate or imply that the device or component referred to must have a specific position, a specific orientation, or be constructed or operated in a specific orientation. Therefore, they should not be construed as limitations on the embodiments of this application.

[0059] Unless otherwise expressly specified or limited, the terms "installation," "connection," "linking," "fixing," and "setting," as used in the description of the embodiments of this application, should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral setting; it can be a mechanical connection, an electrical connection, or a communication connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal connection of two components or the interaction between two components. For those skilled in the art to which this application pertains, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0060] like Figure 1 The diagram shown illustrates the distribution of the power rollers involved in the vacuum coating process parameter adaptive adjustment system of this embodiment. It includes 11 power rollers, arranged as unwinding roller F1, flattening rollers F2-F5, partition roller F6, and traction rollers F7-F8. 10, take-up roller F 11 The rollers are arranged in sequence along the base film transport path, with the separator roller F6 being a 300mm diameter rubber roller with a 120° wrap angle, and the traction rollers F7-F... 10 These are metal rollers with built-in cooling channels. The 11 power rollers of the main production line are fixed by stainless steel brackets, including the unwinding roller F1 and the take-up roller F... 11 It adopts an air-expanding shaft structure, suitable for rolls with diameters of 500~1500mm; the F6 partition roller is coated with polyurethane material with a Shore hardness of 85° to ensure non-slip transmission of the base film; traction rollers F7-F 10 The internal coolant is -5℃ and the cooling circulation is achieved through a rotary joint.

[0061] like Figure 2 As shown, in this embodiment, the adaptive adjustment system for vacuum coating process parameters includes:

[0062] The main body of the vacuum coating production line has multiple power rollers arranged sequentially along the base film transport path, including unwinding rollers, flattening rollers, partition rollers, traction rollers and take-up rollers.

[0063] The data acquisition layer includes tension sensors installed between each roller, a sheet resistance detector for detecting film thickness, a CCD vision system for identifying surface defects, and temperature and humidity sensors and vacuum sensors for acquiring environmental parameters of the vacuum chamber.

[0064] The decision layer includes a DRL agent, which comprises a policy network and an evaluation network, and its input is a state vector S fused from multimodal data. t The output is the process parameter adjustment action vector A. t The DRL agent employs a dual-loop learning mechanism, including an offline imitation learning module and an online reinforcement learning module. The offline imitation learning module is configured to: process the expert dataset D... expert Cleaning, normalization, and alignment preprocessing are performed; the policy network π is trained using the behavior cloning algorithm. θ In expert state S t For input, expert action A t For the label, minimize the mean squared error loss function L BC (θ)=E[(π θ (S t ), A t )²], output the initial network parameters θ init .

[0065] The execution layer includes a servo motor controller connected to each power roller and a power controller connected to the evaporation source. The servo motor controller supports synchronous operation mode and speed loop saturation torque limiting mode.

[0066] A knowledge base is used to store expert datasets D, which consist of historical high-quality process data. expert The dataset contains state-action-reward tuples (S... t A t ,R t ).

[0067] The construction of the expert dataset includes:

[0068] Collect historical high-quality process data; each record is: D expert =(S t A t ,R t ), ;

[0069] A t =[ΔT set ,ΔV roll ,ΔP evap [R represents expert workload] t For subsequent membrane quality scoring. (T) set For the winding and unwinding tension, the main linear speed is V. roll The evaporation source power is P evap .

[0070] Network pre-training:

[0071] Policy Network π θ Learning expert strategies through behavioral cloning: ;

[0072] After initialization, the network has basic control capabilities.

[0073] The data acquisition layer is equipped with tension sensors with an accuracy of ±1N (one for each of the 11 rollers), a sheet resistance detector with a resolution of 0.1Ω / , a CCD vision system with a frame rate of 30fps (the CCD vision system is equipped with a YOLOv5 model, the training dataset contains 10,000 labeled images, and the average accuracy is 92.3%), and a measurement range of -20~80℃, 0~100%RH, (10 -5 Temperature, humidity, and vacuum sensors (~10 Pa). In the data acquisition layer, the tension sensor is an S-type tension sensor, installed on the tension detection frame between each roller, with a contact pressure of <0.5N with the base film; the sheet resistance detector uses the four-probe method, with the detection head 1m away from the evaporation source outlet, outputting sheet resistance data in real time, and using the formula H=k / ρ (k is the material coefficient, ρ is the sheet resistance) to back-calculate the film thickness; the CCD vision system contains two industrial cameras, which respectively capture images of the upper and lower surfaces of the base film, and transmit the images to the decision layer in real time via Ethernet; the environmental sensor is installed on the inner wall of the vacuum chamber.

[0074] The execution layer employs a servo motor controller (response time <10ms) supporting synchronous operation mode and an evaporator source power controller with power regulation accuracy of ±10W; the knowledge base is built on an SQL Server database, storing high-quality process data from 500 batches of expert operations, with each data entry containing a status-action-reward tuple (S... t A t ,R t ).

[0075] The data acquisition layer synchronously acquires multi-source data periodically (e.g., every 200ms): the tension sensor outputs the real-time tension value F between 11 rollers. 1..11 The sheet resistance detector acquires the sheet resistance of the coating and converts it into film thickness data; the CCD vision system uses the YOLOv5 model to identify wrinkled and yellowed areas, and calculates the wrinkled area ratio R. f (Pixel ratio) and yellowing level N s (0-1 normalization, 0 for no color difference, 1 for severe yellowing); the environmental sensor collects the vacuum chamber temperature T. V Humidity H V Vacuum degree P V All data is transmitted to the state coding module of the decision-making layer after A / D conversion.

[0076] In the execution layer, the servo motor controller receives speed and tension commands from the decision layer via the PROFINET bus, controlling the unwinding roll F1 and the take-up roll F2. 11 Operating in speed loop saturation torque limiting mode, the remaining power rollers operate in speed mode, and the acceleration and deceleration times of all motors are uniformly set to 0.5s to ensure synchronous operation; the evaporation source power controller adopts PID+PWM regulation mode, responding to the decision layer's ΔP evap instruction.

[0077] The knowledge base adopts an iterative mechanism of "collection-filtering-updating":

[0078] Data collection: After each batch of production is completed, process parameters, test data and reward values ​​are automatically archived;

[0079] Data filtering: High-quality data with a reward value > 0.8 are retained and added to the expert dataset D. expert ;

[0080] Model update: Offline imitation learning retraining is triggered monthly to update the initial parameters θ of the policy network. init This allows the initial control performance to continuously improve as data accumulates.

[0081] In this embodiment, the inter-roller tension collaborative control algorithm is configured in the decision layer, and the algorithm includes:

[0082] Tension transfer model in the unwinding zone: T5 = F1 - ΣFi , i=2 to 5, where T5 is the tension at the input end of the partition roller, F1 is the tension of the unwinding roller, and F2-F5 are the tensions of the flattening roller;

[0083] Tension transfer model in the winding zone: T6=F 11 +ΣF i (i=7 to 10), where T6 is the tension at the output end of the partition roller, F 11 For the tension of the take-up roll, F7-F 10 For the tension of the traction roller;

[0084] Speed ​​gain adjustment rules: Speed ​​gain ΔG of each power roller i Linear distribution by location:

[0085] , , where i mid The position of the partition roller is given by β, and β is the gain coefficient. β is calibrated using feedback from the tension sensor to ensure that |T5'-T5| < 5N. A negative gain is used in the unwinding zone, a positive gain in the winding zone, and the speed gain of the partition roller is 0. The gain coefficient β is calibrated using feedback from the tension sensor to ensure that the absolute value of the error between the measured tension and the model calculation is less than a preset value.

[0086] In this embodiment, the preprocessing module integrated into the DRL agent includes a tension transfer model for the unwinding and rewinding zones and speed gain adjustment logic:

[0087] Tension transfer model in the unwinding zone: T5 is the tension at the input end of the partition roller F6, and F1-F5 are obtained in real time by tension sensors.

[0088] Tension transfer model in the winding zone: Where T6 is the tension at the output end of the partition roller F6, and F7-F 11 This is the real-time detection value.

[0089] Speed ​​gain adjustment rules: The preset gain gradient for the unwinding zone is [-4‰, -3‰, -2‰, -1‰] (F2-F5), and the gain gradient for the take-up zone is [1‰, 2‰, 3‰, 4‰] (F7-F5). 10 The gain of the partition roller F6 is 0, and it is dynamically calibrated by the gain coefficient β (illustrative, the initial value of β is 1.0, and it is updated every 100ms according to the tension error: β=β-0.01×(T measured-T calculated)).

[0090] After film threading, the take-up and unwind rolls adopt a speed loop saturated torque limiting mode (take-up speed set to 1.05 times the main line speed, unwind speed set to 0.95 times the main line speed) to establish initial tension (e.g., 200N). When the main line runs at low speed (5m / min), the algorithm calculates T5 and T6 through the tension transmission model and compares them with the measured values ​​T5' and T6' of the tension sensor: if |T5'-T5|≤5N and |T6'-T6|≤5N, the model is considered to be suitable; if the error exceeds the threshold, the mechanical transmission mechanism is calibrated (checking bearing friction and reduction ratio deviation). This algorithm achieves decoupled control of multi-roll tension through linear distribution of speed gain and real-time calibration, reducing the tension fluctuation of ±10% in traditional PID to ±3%, and solving the stress concentration and wrinkling problems caused by multi-roll coupling.

[0091] like Figure 4 As shown, the online reinforcement learning module is configured as follows:

[0092] Real-time status data is collected periodically, and action A is output through the initialized policy network. t And execute;

[0093] Transfer sample (S) t A t ,R t ,S t-1 Stored in a FIFO-structured experience replay pool;

[0094] Batch data is randomly sampled from the replay pool periodically. The evaluation network parameter φ is updated to minimize the Bellman error, the policy network parameter θ is updated to maximize the expected return, and the target network is updated through soft update rules θ'←τθ+(1-τ)θ' and φ'←τφ+(1-τ)φ', where τ is the soft update coefficient.

[0095] The state vector S t =[F 1..11 D expert ᵣ,T m ,R f N s ,T v H v ,P v ], where F 1..11 For the tension values ​​of 11 power rollers, D r T is the roller diameter parameter. m R is the film surface temperature. f N represents the percentage of the folded area. s T represents the degree of yellowing of the coating. v H v P v These are the vacuum chamber temperature, humidity, and vacuum level, respectively.

[0096] In this embodiment, the implementation of the offline imitation learning module includes: constructing a policy network π based on the PyTorch framework. θ (A 3-layer fully connected neural network, 8 input dimensions, 256 hidden neurons, 3 output dimensions) and the evaluation network Qᵩ (structured similarly to the policy network). Expert dataset D expert Preprocessing steps:

[0097] Data cleaning: Remove abnormal data with tension fluctuation > ±15% and defect rate > 10%;

[0098] Normalization: Map parameters such as tension (0-500N) and temperature (-20-80℃) to the [0,1] range;

[0099] Alignment processing: Interpolation is performed at a sampling period of 200ms to ensure that the state and action timing match.

[0100] Behavioral cloning training process:

[0101] Loss function:

[0102] Optimizer: Adam optimizer, learning rate 0.001;

[0103] Training iterations: Batch size 64, 1000 iterations until the loss function converges (<0.01), output initial parameters θ. init .

[0104] When from expert dataset D expert A large number of state-action pairs (S) are randomly selected from the data. t A t When calculating the action π predicted by the neural network, θ (S t ) and expert real action A t The difference of squares between ||(π) θ (S t )- A t The average value of ||² is given by the policy network π. θ A neural network with parameters θ, input: the state S at a certain time. t Output: Predicted action π θ (S t In short: L BC The loss function (θ) calculates the average squared error between the action predicted by the neural network and the action actually performed by the expert. The smaller the loss value, the better the neural network imitates the expert.

[0105] The online reinforcement learning module implementation includes:

[0106] State quantity description in the state space:

[0107] F{ 1..11} represents the tension value (N) of 11 rollers;

[0108] T m The film surface temperature (°C) is the temperature of the membrane.

[0109] R f The percentage of wrinkled area (%)

[0110] N s The degree of yellowing of the coating (normalized from 0 to 1);

[0111] T v Vacuum cavity temperature (°C), H v Humidity (%), P v Vacuum level (Pa);

[0112] Action space: ;

[0113] State acquisition: Acquire state vector every 200ms.

[0114] ;

[0115] Where D r T represents the measured roller diameter (obtained using a laser diameter gauge). m The membrane surface temperature (detected by an infrared thermometer);

[0116] Action output: The policy network outputs actions based on θ. ;

[0117] Where ΔT set The tension adjustment amount (±20N), ΔV roll The main line speed adjustment (±0.5m / min), ΔP evap This is the evaporation power adjustment amount (±50W);

[0118] Experience storage: transferring samples Storage capacity 10 6 FIFO experience replay pool;

[0119] Network update: An update is performed every 60 seconds.

[0120] Sampling: 256 samples were randomly selected from the playback pool;

[0121] Critic Update: Calculate the target Q value Minimize the loss function ;

[0122] Actor Update: Through Policy Gradient:

[0123] Maximize expected return.

[0124] Soft update: .

[0125] As shown in Table 1 below:

[0126]

[0127] Table 1

[0128] In Table 1, 1.● represents a sampling point, and the state S is collected every 200ms. t ★ indicates an update operation;

[0129] 2. Experience storage format: (state S) t Action A t Reward R t The next state S{ t+1});

[0130] 3. The network update cycle is 60,000ms (1 minute), and four steps are performed during the update;

[0131] 4. Soft update formula: φ'←τφ+(1-τ)φ', where τ is the update coefficient (0<τ<<1).

[0132] In this embodiment, the dual-loop mechanism addresses the cold start and dynamic adaptation issues in DRL industrial applications through a collaborative "offline pre-training - online optimization" model: offline imitation learning enables the initial strategy to reach expert-level operation (tension fluctuation ±5%), avoiding increased scrap rates due to trial and error; online learning continuously optimizes network parameters through real-time interaction with the production environment, achieving a compensation response time of <1s for disturbances such as roller diameter thermal deformation (e.g., increasing F7 roller diameter from 200mm to 200.2mm) and vacuum fluctuations. Actual measurements show that this mechanism reduces reinforcement learning convergence time by more than 50% and significantly improves film thickness uniformity under dynamic operating conditions.

[0133] In this embodiment, the decision layer further includes a reward function module, wherein the reward function is a multi-objective weighted reward function:

[0134] Where ΔF is the change in the standard deviation of tension (to suppress fluctuations), ΔH is the deviation between the measured film thickness and the target thickness, and R... f I represents the percentage of the folded area. stable For stability bonus (10 consecutive steps without overshoot = 1), the weighting coefficients are w1=0.3, w2=0.4, w3=0.2, and w4=0.1.

[0135] This embodiment also includes a safety constraint module, which is configured as follows:

[0136] Action correction layer: Action A output by DRL t Physically limit the tension adjustment to ensure it does not exceed 10% of the current setting.

[0137] Emergency intervention unit: When the percentage of wrinkle area R is detected f When the thickness deviation is >20% or the film thickness deviation |ΔH|>15%, the expert system will take over control.

[0138] The state coding module can be implemented using the Python numpy library, which merges the eight physical quantities into a 1×8 dimensional vector:

[0139] Tension component: The mean and standard deviation of the tension of 11 rollers are taken and encoded as two features;

[0140] Roller diameter and film temperature: directly take D r With T m The normalized value;

[0141] Visual component: R f (Percentage normalization) and N s (Enter directly);

[0142] Environmental component: T v H v P v The normalized value.

[0143] The reward function is defined by the formula:

[0144] Implementation; where:

[0145] ΔF: The difference in standard deviation of tension between the current cycle and the previous cycle; ΔH: The absolute value of the deviation of the measured film thickness from 1000 nm; I stable The value is 1 if there is no overshoot for 10 consecutive steps (tension fluctuation < ±3%, film thickness deviation < ±5%), otherwise it is 0; the weighting coefficients are fixed at w1=0.3, w2=0.4, w3=0.2, w4=0.1, and can be manually adjusted via an industrial PC HMI.

[0146] In this embodiment, multimodal state fusion enables the DRL agent to simultaneously "sense" the equipment's operating status (tension, temperature), product quality status (wrinkles, yellowing), and environmental interference (vacuum level), breaking through the limitations of traditional control methods that rely solely on physical quantities. The reward function, through weighted multi-objective optimization, guides the agent to prioritize suppressing wrinkle defects and tension fluctuations, balancing film thickness accuracy and operational stability, thereby significantly reducing the wrinkle defect rate and achieving a balance between quality and efficiency. This module addresses potential dangerous actions during DRL exploration: action limiting prevents film breakage caused by sudden tension changes, and emergency intervention quickly stops losses when serious defects occur.

[0147] like Figure 3 As shown, in another embodiment, a method for adaptively adjusting vacuum coating process parameters based on the system described in the above embodiments is provided. This method includes the following steps:

[0148] Initialize process parameters: Based on the input base film material, width, and target film thickness, match the initial formula from the knowledge base or generate it through expert rules. The formula includes the winding and unwinding tension T. set Main line velocity V roll and evaporation source power P evap ;

[0149] Inter-roll tension coordination calibration: After film threading, the winding and unwinding tension is established. The tension on both sides of the partition roller is calculated through the tension transmission model. The model is calibrated by comparing the actual sensor values ​​and adjusting the speed gain of each roller to balance the tension of each section of film.

[0150] Vacuum environment adaptation: After the base film enters the vacuum chamber, the main line runs at low speed, and the speed gain of each roller is finely adjusted to compensate for the tension imbalance caused by the change in the physical properties of the roller;

[0151] DRL Dual-Cycle Optimization: After the vacuum level and evaporation power meet the standards, the main line runs at high speed and deposits a film. The DRL agent collects multi-modal status in real time and outputs action to adjust process parameters after dual-cycle learning.

[0152] Safety monitoring and data archiving: During the production process, the operating status is monitored through the safety constraint module, process parameters are stored in the database in real time, and the formula is added to the knowledge base after production is completed.

[0153] The adaptive adjustment method for vacuum coating process parameters further includes the following steps:

[0154] When the percentage of wrinkle area R is detected f When the thickness deviation is >20% or the film thickness deviation |ΔH|>15%, the expert system will take over control.

[0155] In step S1,

[0156] Input product attributes: Base film material PET, width 1.8m, target film thickness 1000nm. Match historical similar formulations from the knowledge base: T set =200N, V roll =10m / min, P evap =5kW. For new specifications, the expert rule applies:

[0157] , Generate initial parameters, where W film Where is the film width, u is the coefficient of friction, and H is... targetWhere W is the target film thickness, and k1 and k2 are material correlation coefficients (k1=0.5, k2=10, W) film (where μ is the film width and μ is the coefficient of friction, 0.3).

[0158] During the inter-roller tension coordination calibration, after film threading, start winding and unwinding to establish an initial tension of 200N, with the main line running at low speed (5m / min). Calculate T5=200N and T6=200N using the tension transfer model, and compare with the actual sensor values ​​T5'=198N and T6'=202N (error ≤5N). No model calibration is required. Adjust the gain of F2-F5 to -4‰~-1‰, and F7-F... 10 The gain is 1‰~4‰, which stabilizes the membrane tension in each segment at 200N±3N.

[0159] During the vacuum environment adaptation process, after the base film enters the vacuum chamber, the chamber door is closed and a vacuum is drawn until... Pa, start the evaporator roller cooling system. After a period of time (e.g., after 10 minutes), it is detected that the F7 roller diameter increases from 200mm to 200.2mm, and the tension fluctuates to 210N. The DRL agent fine-tunes the F7 gain to 0.2‰, and the tension recovers to 200N±3N.

[0160] After the vacuum level and evaporation power meet the standards, the main line speed is increased to 10 m / min, and the evaporation source baffle and bias power supply (voltage 500V) are turned on. The CCD vision system detects R. f =3%, N s =0.1, sheet resistance detection shows film thickness 980nm, DRL agent output action: ΔT set =+5N、ΔV roll =-0.2m / min, ΔP evap =+30W. After execution, the film thickness recovered to 995nm, R f Reduced to 1.5%. During continuous operation, the system updates its status every 200ms, optimizes the network every 60 seconds, and compensates for environmental disturbances in real time.

[0161] In this embodiment, the inter-roller tension collaborative control algorithm is configured in the decision layer, and the algorithm includes:

[0162] Tension transfer model in the unwinding zone: T5 = F1 - ΣF i , i=2 to 5, where T5 is the tension at the input end of the partition roller, F1 is the tension of the unwinding roller, and F2-F5 are the tensions of the flattening roller;

[0163] Tension transfer model in the winding zone: T6=F 11 +ΣF i (i=7 to 10), where T6 is the tension at the output end of the partition roller, F 11 For the tension of the take-up roll, F7-F 10For the tension of the traction roller;

[0164] Speed ​​gain adjustment rules: Speed ​​gain ΔG of each power roller i The unwinding zone is linearly distributed according to position, and the winding zone uses negative gain. The speed gain of the break roller is 0. The gain coefficient β is calibrated by the feedback value of the tension sensor so that the absolute value of the error between the measured tension and the model calculation value is less than the preset value.

[0165] Finally, it should be noted that although the above embodiments have been described in the text and drawings of this application, this should not limit the scope of patent protection of this application. Any technical solutions that are based on the essential concept of this application and utilize the content described in the text and drawings of this application, resulting in equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this application.

Claims

1. An adaptive adjustment system for vacuum coating process parameters, characterized in that, include: The main body of the vacuum coating production line has multiple power rollers arranged sequentially along the base film transport path, including unwinding rollers, flattening rollers, dividing rollers, traction rollers and take-up rollers; The data acquisition layer includes tension sensors installed between each roller, a sheet resistance detector for detecting film thickness, a CCD vision system for identifying surface defects, and temperature and humidity sensors and vacuum sensors for acquiring environmental parameters of the vacuum chamber. The decision layer includes a DRL agent, which comprises a policy network and an evaluation network, and its input is a state vector S fused from multimodal data. t The output is the process parameter adjustment action vector A. t The DRL agent employs a dual-loop learning mechanism, including an offline imitation learning module and an online reinforcement learning module. The offline imitation learning module is configured to: process the expert dataset D... expert Cleaning, normalization, and alignment preprocessing are performed; the policy network π is trained using the behavior cloning algorithm. θ In expert state S t The input is AT, the expert action is AT, and the loss function L is minimized. BC (θ)=E[(π θ (S t A t )²], output the initial network parameters θ init ; The execution layer includes a servo motor controller connected to each power roller and a power controller connected to the evaporation source. The servo motor controller supports synchronous operation mode and speed loop saturation torque limiting mode. A knowledge base is used to store expert datasets D, which consist of historical high-quality process data. expert The dataset contains state-action-reward tuples (S... t A t ,R t ).

2. The adaptive adjustment system for vacuum coating process parameters according to claim 1, characterized in that, The inter-roller tension collaborative control algorithm is configured in the decision layer, and the algorithm includes: Tension transfer model in the unwinding zone: T5 = F1 - ΣF i , i=2 to 5, where T5 is the tension at the input end of the partition roller, F1 is the tension of the unwinding roller, and F2-F5 are the tensions of the flattening roller; Tension transfer model in the winding zone: T6=F 11 +ΣF i (i=7 to 10), where T6 is the tension at the output end of the partition roller, F 11 For the tension of the take-up roll, F7-F 10 For the tension of the traction roller; Speed ​​gain adjustment rules: Speed ​​gain ΔG of each power roller i The unwinding zone is linearly distributed according to position, and the winding zone uses negative gain. The speed gain of the break roller is 0. The gain coefficient β is calibrated by the feedback value of the tension sensor so that the absolute value of the error between the measured tension and the model calculation value is less than the preset value.

3. The adaptive adjustment system for vacuum coating process parameters according to claim 1, characterized in that, The online reinforcement learning module is configured as follows: Real-time status data is collected periodically, and action A is output through the initialized policy network. t And execute; Transfer sample (S) t A t ,R t ,S t+1 Stored in a FIFO-structured experience replay pool; Batch data is randomly sampled from the replay pool periodically. The evaluation network parameter φ is updated to minimize the Bellman error, the policy network parameter θ is updated to maximize the expected return, and the target network is updated through soft update rules θ'←τθ+(1-τ)θ' and φ'←τφ+(1-τ)φ', where τ is the soft update coefficient.

4. The adaptive adjustment system for vacuum coating process parameters according to claim 1, characterized in that, The state vector S t =[F 1..11 D expert ᵣ,T m ,R f N s ,T v H v ,P v ], where F 1..11 For the tension values ​​of 11 power rollers, D r T is the roller diameter parameter. m R is the film surface temperature. f N represents the percentage of the folded area. s T represents the degree of yellowing of the coating. v H v P v These are the vacuum chamber temperature, humidity, and vacuum level, respectively.

5. The adaptive adjustment system for vacuum coating process parameters according to claim 1, characterized in that, The decision-making layer also includes a reward function module, wherein the reward function is a multi-objective weighted reward function: ΔF is the change in the standard deviation of tension, ΔH is the deviation between the measured film thickness and the target thickness, and R... f I represents the percentage of the folded area. stable As a stability bonus, the weighting coefficients are w1=0.3, w2=0.4, w3=0.2, and w4=0.

1.

6. The adaptive adjustment system for vacuum coating process parameters according to claim 1, characterized in that, It also includes a safety constraint module, which is configured as follows: Action correction layer: Action A output by DRL t Physically limit the tension adjustment to ensure it does not exceed 10% of the current setting. Emergency intervention unit: When the percentage of wrinkle area R is detected f When the thickness deviation is >20% or the film thickness deviation |ΔH|>15%, the expert system will take over control.

7. A method for adaptive adjustment of vacuum coating process parameters based on the system described in any one of claims 1-6, characterized in that, Includes the following steps: S1. Initialize process parameters: Based on the input base film material, width, and target film thickness, match the initial formula from the knowledge base or generate it through expert rules. The formula includes the winding and unwinding tension T. set Main line velocity V roll and evaporation source power P evap ; S2. Coordinated calibration of inter-roller tension: After film insertion, the winding and unwinding tension is established. The tension on both sides of the partition roller is calculated through the tension transmission model. The model is calibrated by comparing the actual sensor values ​​and adjusting the speed gain of each roller to balance the tension of each section of film. S3. Vacuum environment adaptation: After the base film enters the vacuum chamber, the main line runs at low speed and the speed gain of each roller is finely adjusted to compensate for the tension imbalance caused by the change in the physical properties of the roller; S4.DRL Dual-Cycle Optimization: After the vacuum level and evaporation power meet the standards, the main line runs at high speed and coats the film. The DRL intelligent agent collects multi-modal status in real time and outputs action to adjust process parameters after dual-cycle learning. S5. Safety monitoring and data archiving: During the production process, the operating status is monitored through the safety constraint module, process parameters are stored in the database in real time, and the formula is added to the knowledge base after production is completed.

8. The adaptive adjustment method for vacuum coating process parameters of the system according to claim 7, characterized in that, The inter-roller tension collaborative control algorithm is configured in the decision layer, and the algorithm includes: Tension transfer model in the unwinding zone: T5 = F1 - ΣF i , i=2 to 5, where T5 is the tension at the input end of the partition roller, F1 is the tension of the unwinding roller, and F2-F5 are the tensions of the flattening roller; Tension transfer model in the winding zone: T6=F 11 +ΣF i (i=7 to 10), where T6 is the tension at the output end of the partition roller, F 11 For the tension of the take-up roll, F7-F 10 For the tension of the traction roller; Speed ​​gain adjustment rules: Speed ​​gain ΔG of each power roller i The unwinding zone is linearly distributed according to position, and the winding zone uses negative gain. The speed gain of the break roller is 0. The gain coefficient β is calibrated by the feedback value of the tension sensor so that the absolute value of the error between the measured tension and the model calculation value is less than the preset value.

9. The adaptive adjustment method for vacuum coating process parameters of the system according to claim 7, characterized in that, It also includes the following steps: When the percentage of wrinkle area R is detected f When the thickness deviation is >20% or the film thickness deviation |ΔH|>15%, the expert system will take over control.

Images